Clustering Execution & Optimization

Resumen

Optimización de hiperparámetros y ejecución de algoritmos de clustering sobre espacios MOFA y UMAP.

Añadir al sistema al ruta al directorio base para importar módulos personalizados.

import sys, os
sys.path.append(os.path.abspath(".."))

Importar librerías y módulos necesarios.

import pandas as pd
import src.data_utils as du
import src.clustering_utils as cu
import src.plots as p
import seaborn as sns
import matplotlib.pyplot as plt
import numpy as np

Importar datos procesados de MOFA y UMAP.

DATA_DIR = "../data/"

# Factores de MOFA
m_factors = du.load_data(DATA_DIR + "MOFA_dir/M_factors", index_col=0)

# Caracteristicas de UMAP
m_umap = du.load_data(DATA_DIR + "processed_data/M_umap", index_col=0)

Ejecución de K-Means

Datos MOFA

Selección de Hyperparámetros óptimos

mofa_kmeans_res = cu.run_kmeans_optimization(m_factors, k_range=range(2, 12))
opt_k_mofa = mofa_kmeans_res["optimal_k"]

print(f"MOFA Optimal k detected: {opt_k_mofa}")

p.plot_elbow_silhouette(
    mofa_kmeans_res["k_range"],
    mofa_kmeans_res["inertias"],
    mofa_kmeans_res["silhouettes"],
    opt_k_mofa,
    "MOFA Clustering",
)

MOFA Optimal k detected: 7

Matriz de proximidad de los clusters

labels_mofa = mofa_kmeans_res["results"][8]["labels"]
m_factors_ordered = m_factors.copy()
m_factors_ordered["Cluster"] = labels_mofa
m_factors_ordered = m_factors_ordered.sort_values("Cluster")

dist_matrix = cu.get_proximity_matrix(m_factors_ordered.drop(columns="Cluster"))

plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by K-Means MOFA Cluster)")
plt.show()

Datos UMAP

Selección de Hyperparámetros óptimos

umap_kmeans_res = cu.run_kmeans_optimization(m_umap, k_range=range(2, 12))
opt_k_umap = umap_kmeans_res["optimal_k"]
selected_k_umap = 8

print(f"UMAP Optimal k detected: {opt_k_umap}")

p.plot_elbow_silhouette(
    umap_kmeans_res["k_range"],
    umap_kmeans_res["inertias"],
    umap_kmeans_res["silhouettes"],
    selected_k_umap,
    "UMAP Clustering",
)

UMAP Optimal k detected: 6

Matriz de proximidad de los clusters

labels_umap = umap_kmeans_res["results"][8]["labels"]
m_umap_ordered = m_umap.copy()
m_umap_ordered["Cluster"] = labels_umap
m_umap_ordered = m_umap_ordered.sort_values("Cluster")

dist_matrix = cu.get_proximity_matrix(m_umap_ordered.drop(columns="Cluster"))

plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by K-Means UMAP Cluster)")
plt.show()

Ejecución de GMM

Datos MOFA

Selección de Hyperparámetros óptimos

gmm_res = cu.optimize_gmm(m_factors, k_range=range(2, 12))
p.plot_gmm_optimization(
    gmm_res["k_range"], gmm_res["bics"], gmm_res["aics"], gmm_res["silhouettes"]
)
best_k_gmm = gmm_res["k_range"][np.argmin(gmm_res["bics"])]
print(f"Best K based on BIC: {best_k_gmm}")

k_gmm_mofa = 8  # Seleccionado manualmente
labels_gmm_opt = gmm_res["results"][k_gmm_mofa]["labels"]

Best K based on BIC: 9

Matriz de proximidad de los clusters

m_factors_ordered_gmm = m_factors.copy()
m_factors_ordered_gmm["Cluster"] = labels_gmm_opt
m_factors_ordered_gmm = m_factors_ordered_gmm.sort_values("Cluster")
dist_matrix_gmm = cu.get_proximity_matrix(
    m_factors_ordered_gmm.drop(columns="Cluster")
)
plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix_gmm, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by GMM MOFA Cluster)")
plt.show()

Datos UMAP

Selección de Hyperparámetros óptimos

gmm_umap_res = cu.optimize_gmm(m_umap, k_range=range(2, 12))
p.plot_gmm_optimization(
    gmm_umap_res["k_range"], gmm_umap_res["bics"], gmm_umap_res["aics"], gmm_umap_res["silhouettes"]
)
best_k_gmm_umap = gmm_umap_res["k_range"][np.argmin(gmm_umap_res["bics"])]
print(f"Best K based on BIC: {best_k_gmm_umap}")
k_gmm_umap = 8  # Seleccionado manualmente
labels_gmm_umap_opt = gmm_umap_res["results"][k_gmm_umap]["labels"]

Best K based on BIC: 8

Matriz de proximidad de los clusters

m_umap_ordered_gmm = m_umap.copy()
m_umap_ordered_gmm["Cluster"] = labels_gmm_umap_opt
m_umap_ordered_gmm = m_umap_ordered_gmm.sort_values("Cluster")
dist_matrix_gmm_umap = cu.get_proximity_matrix(
    m_umap_ordered_gmm.drop(columns="Cluster")
)
plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix_gmm_umap, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by GMM UMAP Cluster)")
plt.show()

Ejecución de DBSCAN

Datos MOFA

Selección de Hyperparámetros óptimos

eps_candidates = np.arange(0.1, 2.0, 0.1)
min_samples_candidates = [3, 5, 10]

dbscan_table = cu.grid_search_dbscan(m_factors, eps_candidates, min_samples_candidates)

dbscan_table_filtered = dbscan_table[
    (dbscan_table["n_clusters"] > 1) & (dbscan_table["noise_ratio"] < 1)
].sort_values("silhouette", ascending=False)

display(dbscan_table_filtered.head(10))

	eps	min_samples	n_clusters	n_noise	noise_ratio	silhouette	labels
39	1.4	3	2	153	0.962	-0.154	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...
42	1.5	3	2	153	0.962	-0.154	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...
54	1.9	3	9	120	0.755	-0.181	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...
45	1.6	3	4	146	0.918	-0.207	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...
48	1.7	3	6	138	0.868	-0.224	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...
51	1.8	3	8	129	0.811	-0.237	[-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -...

k_neighbors = 5
distances = cu.calculate_k_distance(m_factors, k=k_neighbors)
p.plot_k_distance_curve(distances, k=k_neighbors)

labels_dbscan_opt = cu.run_dbscan(m_factors, eps=4, min_samples=5)["labels"]

Matriz de proximidad de los clusters

m_factors_ordered_dbscan = m_factors.copy()
m_factors_ordered_dbscan["Cluster"] = labels_dbscan_opt
m_factors_ordered_dbscan = m_factors_ordered_dbscan.sort_values("Cluster")
dist_matrix_dbscan = cu.get_proximity_matrix(
    m_factors_ordered_dbscan.drop(columns="Cluster")
)
plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix_dbscan, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by DBSCAN MOFA Cluster)")
plt.show()

Datos UMAP

Selección de Hyperparámetros óptimos

eps_candidates = np.arange(0.1, 2.0, 0.1)
min_samples_candidates = [3, 5, 10]
dbscan_umap_table = cu.grid_search_dbscan(m_umap, eps_candidates, min_samples_candidates)
dbscan_umap_table_filtered = dbscan_umap_table[
    (dbscan_umap_table["n_clusters"] > 1) & (dbscan_umap_table["noise_ratio"] < 1)
].sort_values("silhouette", ascending=False)
display(dbscan_umap_table_filtered.head(10))

	eps	min_samples	n_clusters	n_noise	noise_ratio	silhouette	labels
22	0.8	5	6	0	0.000	0.608	[0, 0, 1, 0, 0, 1, 1, 2, 1, 1, 2, 2, 0, 1, 1, ...
18	0.7	3	6	0	0.000	0.608	[0, 0, 1, 0, 0, 1, 1, 2, 1, 1, 2, 2, 0, 1, 1, ...
23	0.8	10	7	1	0.006	0.594	[0, 0, 2, 0, 0, 2, 2, 1, 2, 2, 1, 1, 0, 2, 3, ...
20	0.7	10	9	8	0.050	0.590	[-1, 0, 2, 0, 0, 2, 2, 1, 2, 2, 1, 1, 3, 2, 4,...
21	0.8	3	5	0	0.000	0.580	[0, 0, 1, 0, 0, 1, 1, 2, 1, 1, 2, 2, 0, 1, 1, ...
19	0.7	5	6	1	0.006	0.570	[0, 0, 1, 0, 0, 1, 1, 2, 1, 1, 2, 2, 0, 1, 1, ...
15	0.6	3	7	0	0.000	0.554	[0, 0, 1, 2, 0, 1, 1, 3, 1, 1, 3, 3, 2, 1, 1, ...
13	0.5	5	10	7	0.044	0.545	[0, 0, 1, 3, 0, 1, 1, 2, 1, -1, 2, 2, 3, 6, 4,...
25	0.9	5	5	0	0.000	0.544	[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, ...
29	1.0	10	5	0	0.000	0.544	[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, ...

k_neighbors = 5
distances_umap = cu.calculate_k_distance(m_umap, k=k_neighbors)
p.plot_k_distance_curve(distances_umap, k=k_neighbors)

labels_dbscan_umap_opt = cu.run_dbscan(m_umap, eps=0.4, min_samples=5)["labels"]

Matriz de proximidad de los clusters

m_umap_ordered_dbscan = m_umap.copy()
m_umap_ordered_dbscan["Cluster"] = labels_dbscan_umap_opt
m_umap_ordered_dbscan = m_umap_ordered_dbscan.sort_values("Cluster")
dist_matrix_dbscan_umap = cu.get_proximity_matrix(
    m_umap_ordered_dbscan.drop(columns="Cluster")
)
plt.figure(figsize=(8, 6))
sns.heatmap(dist_matrix_dbscan_umap, cmap="viridis", xticklabels=False, yticklabels=False)
plt.title("Proximity Matrix (Sorted by DBSCAN UMAP Cluster)")
plt.show()

Guardar mejores modelos de clustering

# OPCIONES DISPONIBLES:
# labels_kmeans_opt / labels_umap_kmeans_opt
# labels_gmm_opt / labels_gmm_umap_opt
# labels_dbscan_opt / labels_dbscan_umap_opt

CLUTERING_DIR = DATA_DIR + "clustering_dir/"
selected_models = {
    "GMM_Opt": labels_gmm_opt,
    "GMM_UMAP_Opt": labels_gmm_umap_opt,
}
for model_name, labels in selected_models.items():
    df_selected = pd.DataFrame(labels, index=m_factors.index, columns=["Cluster"])
    du.save_data(df_selected, CLUTERING_DIR + f"selected_clusters_{model_name}")

---